We started with finishing the part on Instrumental variables.

Then we moved to studying various research designs: Regression Discontinuity, Difference-in-differences and Synthetic Controls method. Last part was on introduction to the Machine learning methods which finished with Causal Machine Learning - utilizing the great predictive properties of ML algorithms for improving the estimation of parameters of interest (this is what economists are typically interested in).

"Nature does not make jumps". Regression discontinuity design utilizes the variation that is induced by some discontinuity in the data, typically induced by some administrative rule/law.

One of the most popular current research designs is arguably Difference-in-differences. It is everywhere. It accounts for approximately one quarter of the 100 most cited papers in economics from 2015-2019. The main assumption that DiD is based on is so called "parallel trends assumption". We make assumption about the behavior of a unit had the treatment/intervention not happened. There is a lot of important contributions that happened in the literature in the past three years! The methods that were being routinely used have some serious drawbacks. We will go through some of these recent developments too, with an emphasis on our intuition.

Many R illustrations can be found here:

Next research design, Synthetic Control Method, is so beautiful and intuitive it made in into Wall Street Journal or Washington Post. The person who brought and develop these ideas is Alberto Abadie from MIT. For a treated unit there is not a proper match we could compare it to. What do we do? We will create a synthetic one that is a weighted average of different untreated units. The very visual nature of this method makes it very popular in industry too and is an area of current active research.

The last thing we will cover are some basics of Machine Learning and then mention some modern literature how ML is making its way to economics. ML methods are built for prediction. In certain situations we don't care much about estimation or about understanding the underlying mechanism as long as it works/predicts well/earns money. Combining ML algorithms with causal models is where a lot of research efforts are currently directed to. It seems like a good idea to get familiar with basics of ML methods as it is spreading everywhere.

Here is an updated R code for this part with some illustrations:

Assignment 3 - deadline is Dec 15.

Please submit your work here: